OctoInformBeta

OpenAI preprint, September 2026 · §4. Adjoints in moment coordinates

The Hessian identity (4.7): The mean under μ\mu of the Hessian energy tr(τD2fτD2f)\mathrm{tr}(\tau D^2f\,\tau D^2f), written as E[(Af)(A−1)f]E[(Af)(A-1)f] and as E[(Af)2]−E[fAf]E[(Af)^2] - E[f\,Af]

The mean Hessian energy in terms of the moment operator.

Hover over a blue or green term in the text, or a node in the graph, to find it in the other. Select it to read its declaration both in Lean and in words.

The Setting

The coordinate space ℝι\mathbb{R}^\iota over a finite index set ι\iota is the real vector space of vectors x=(xi)i∈ιx = (x_i)_{i \in \iota}, added and scaled coordinatewise.✦ Fix a reference measure on ℝι\mathbb{R}^\iota, written dzdz in source coordinates zz and dxdx in target coordinates xx. For a function φ\varphi, a potential VV and a constant Z>0Z > 0, the source measure ν\nu has density the weight e−φe^{-\varphi}✦, and the target measure μ\mu has density the weight e−Ve^{-V} divided by ZZ: dν(z)=e−φ(z)dz,dμ(x)=1Ze−V(x)dx.d\nu(z) = e^{-\varphi(z)}\,dz, \qquad d\mu(x) = \frac{1}{Z}\,e^{-V(x)}\,dx.(1)✦

The item takes the gradient map ∇φ\nabla\varphi to be a homeomorphism of ℝι\mathbb{R}^\iota that pushes ν\nu forward to μ\mu; when φ\varphi is convex and the reference measure is Lebesgue measure, μ\mu is then the moment measure of φ\varphi. Its subject is a function ff of the target point x=∇φ(z)x = \nabla\varphi(z): the mean under μ\mu of a Hessian energy of ff, built from D2fD^2f and D2φD^2\varphi, written through a second-order operator AA acting on ff in target coordinates.

The Operator and the Energy

The standard basis vector eie_i of ℝι\mathbb{R}^\iota is the vector with coordinate 11 at the index ii and 00 at every other index.✦ The directional derivative DvfD_vf of a function ff on ℝι\mathbb{R}^\iota along a vector vv is the function x↦Df(x)vx \mapsto Df(x)\,v, the Fréchet derivative of ff at xx applied to vv✦, and the partial derivative ∂if\partial_if is DeifD_{e_i}f✦. The gradient ∇f(x)\nabla f(x) of ff at xx is the vector with coordinates ∂if(x)\partial_if(x).✦ The Hessian D2f(x)D^2f(x) is the real matrix indexed by ι×ι\iota \times \iota whose entry in row ii and column jj is ∂j∂if(x)\partial_j\partial_if(x).✦

The expectation E[F]E[F] of a function FF on ℝι\mathbb{R}^\iota under μ\mu is its integral against the weight e−Ve^{-V}, divided by ZZ: E[F]=1Z∫F(x)e−V(x)dx=∫Fdμ.E[F] = \frac{1}{Z}\int F(x)\,e^{-V(x)}\,dx = \int F\,d\mu.(2)✦

Given a homeomorphism Φ\Phi of ℝι\mathbb{R}^\iota and a function φ\varphi on it, the moment matrix τ(x)\tau(x) at a target point xx is the Hessian of φ\varphi at the preimage of xx under Φ\Phi: τ(x)=D2φ(Φ−1(x)).\tau(x) = D^2\varphi\big(\Phi^{-1}(x)\big).(3)✦

For a field τ\tau of real matrices indexed by ι×ι\iota \times \iota, with τ:D2f=∑i,jτij∂i∂jf\tau : D^2f = \sum_{i,j} \tau_{ij}\,\partial_i\partial_jf the Frobenius inner product of τ\tau and D2fD^2f, the operator AA sends a function ff on ℝι\mathbb{R}^\iota to the function Af(x)=∑ixi∂if(x)−∑i,jτij(x)∂i∂jf(x)=x⋅∇f(x)−τ(x):D2f(x).Af(x) = \sum_{i} x_i\,\partial_if(x) - \sum_{i,j} \tau_{ij}(x)\,\partial_i\partial_jf(x) = x \cdot \nabla f(x) - \tau(x) : D^2f(x).(4) When τ\tau is the moment matrix, AA is the moment operator; (A−1)f(A-1)f stands for Af−fAf - f.✦

The Hessian energy of a function ff with respect to a matrix field τ\tau is the function x↦tr(τ(x)D2f(x)τ(x)D2f(x)).x \mapsto \mathrm{tr}\big(\tau(x)\,D^2f(x)\,\tau(x)\,D^2f(x)\big).(5)✦ For the moment matrix of a smooth convex φ\varphi and a smooth ff, it is the squared Hilbert-Schmidt norm of τ1/2D2fτ1/2\tau^{1/2}D^2f\,\tau^{1/2}.

The Claim

Theorem 1. Let φ\varphi and VV be smooth functions on ℝι\mathbb{R}^\iota and Z>0Z > 0, and suppose that the gradient map ∇φ\nabla\varphi is a homeomorphism of ℝι\mathbb{R}^\iota pushing the source measure ν\nu forward to the target measure μ\mu of (1). Let τ\tau be the moment matrix (3) with Φ=∇φ\Phi = \nabla\varphi, AA the moment operator and EE the expectation under μ\mu. Then every smooth compactly supported function ff on ℝι\mathbb{R}^\iota satisfies E[(Af)(A−1)f]=E[tr(τD2fτD2f)].E\big[(Af)\,(A-1)f\big] = E\big[\mathrm{tr}\big(\tau\,D^2f\,\tau\,D^2f\big)\big].(6)✦ If moreover the inverse of ∇φ\nabla\varphi is smooth and μ\mu has finite total mass, then every such ff satisfies E[tr(τD2fτD2f)]=E[(Af)2]−E[fAf].E\big[\mathrm{tr}\big(\tau\,D^2f\,\tau\,D^2f\big)\big] = E\big[(Af)^2\big] - E[f\,Af].(7)✦

What Is Checked

The claim is stated over the source measure ν\nu and the target measure μ\mu with their weights e−φe^{-\varphi} and e−Ve^{-V}, the expectation EE under μ\mu, the moment matrix τ\tau, the operator AA and the Hessian energy, all defined through partial derivatives on the coordinate space ℝι\mathbb{R}^\iota.